Identifying and forecasting importation and asymptomatic spreaders of multi-drug resistant organisms in hospital settings

Jiaming Cui, Jack Heavey, Eili Y. Klein, Gregory Madden, Costi D. Sifri, Anil Vullikanti, B. Aditya PrakashNpj Digit. Medicine 2025 · DOI

Abstract

Healthcare-associated infections (HAIs) from multi-drug resistant organisms (MDROs) pose a significant challenge for healthcare systems. Patients can arrive at hospitals already infected ("importation") or acquire infections during their stay ("nosocomial infection"). Many cases, often asymptomatic, complicate rapid identification due to testing limitations and delays. Although recent advancements in mathematical modeling and machine learning have aimed to identify at-risk patients, these methods face challenges: transmission models often overlook valuable electronic health record (EHR) data, while machine learning approaches typically lack mechanistic insights into underlying processes. To address these issues, we propose NeurABM, a novel framework that integrates neural networks and agent-based models (ABM) to leverage the strengths of both methods. NeurABM simultaneously learns a neural network for patient-level importation predictions and an ABM for infection identification. Our findings show that NeurABM significantly outperforms existing methods, marking a breakthrough in accurately identifying importation cases and forecasting future nosocomial infections in clinical practice.

Introduction

Healthcare-associated infections (HAIs) caused by pathogens like Methicillin-resistant Staphylococcus aureus (MRSA) are a primary driver of hospital mortality and increased healthcare costs. Managing these "superbugs" is difficult because of two primary factors: silent importation and complex nosocomial spread. Many patients arrive at the hospital already colonized but asymptomatic, acting as "silent spreaders." Once inside, the pathogen moves through intricate contact networks involving patients, staff, and the physical environment.

Existing predictive solutions generally fall into two categories: Machine Learning (ML) and Agent-Based Models (ABM). ML models excel at identifying patterns within Electronic Health Records (EHR) but lack mechanistic insight into disease transmission. Conversely, ABMs simulate disease spread effectively but struggle to integrate granular, individual-level patient data and often rely on static assumptions about future contact networks.

Methods: The NeurABM Framework

To bridge this gap, we introduce NeurABM, a hybrid framework that couples a neural network with a differentiable ABM in a single, end-to-end trainable pipeline. This integration allows the system to leverage the predictive power of neural networks for individual risk assessment while maintaining the mechanistic logic of epidemiological simulations.

The framework consists of three primary components:

  1. The Neural Network ($\phi$): This component processes 19 patient-specific risk factors from EHR data (e.g., age, recent surgeries, device usage) to estimate the probability of importation. It also learns general ABM parameters such as transmission and recovery rates.
  2. The Differentiable ABM: Unlike standard simulations, this ABM is differentiable, allowing error signals from the simulation to backpropagate and tune the neural network. It utilizes co-location data to construct contact networks and simulates transitions between "Susceptible" and "Carriage" states.
  3. The Adapter Network ($\phi'$): To address the uncertainty of future contact networks during forecasting, an "adapter" network learns to correct biases introduced by the assumption that future contacts will mirror current ones.

Datasets and Experimental Setup

The study utilized real-world EHR data from the University of Virginia (UVA) hospital intensive care units (ICUs) collected between July and December 2019. The dataset included 1,117 patients, 2,385 healthcare workers, and 77 distinct locations. MRSA infections were identified via laboratory tests (culture and PCR). The model incorporated 19 risk factors, including prior healthcare facility admission, recent MRSA contacts, and length of stay.

Key Results and Insights

NeurABM consistently outperformed all baselines, including XGBoost, standard Feedforward Neural Networks (FNN), and standalone ABMs. The synergy between the neural network and the ABM proved critical: the NN provides the ABM with accurate initial conditions, while the ABM provides a structured environment for the NN to understand infection evolution.

Performance Metrics

  • Identifying Importation: NeurABM achieved an Area Under the Precision-Recall Curve (AUPRC) of 0.60, nearly doubling the performance of the best ML baseline (0.36).
  • Forecasting Future Cases: In forecasting nosocomial infections, NeurABM reached an AUPRC of 0.85, compared to 0.62 for the best ML baseline.
  • Clinical Utility: If a hospital screened only the top 20% of patients flagged as high-risk by NeurABM, they could identify 88% of all MRSA carriage patients. In contrast, the best existing baseline (XGBoost) caught only 65% under the same constraints.
  • Negative Predictive Value (NPV): The model maintained an NPV above 0.9, indicating high reliability when the model classifies a patient as non-carrying.

Discussion and Limitations

While NeurABM represents a significant advancement, several limitations remain. The model currently uses a simplified binary "Susceptible vs. Carriage" state, whereas real-world MRSA involves various clinical infection types with different transmission risks. Additionally, the "well-mixed" assumption—that co-location equals interaction—may be sub-optimal in complex spatial settings. Future work will focus on generalizing the model to other hospital environments and developing more sophisticated methods for forecasting future contact networks.

Summary updated

BibTeX
@article{cui2025identifying,
  author = {Jiaming Cui and Jack Heavey and Eili Y. Klein and Gregory Madden and Costi D. Sifri and Anil Vullikanti and B. Aditya Prakash},
  title = {{Identifying and forecasting importation and asymptomatic spreaders of multi-drug resistant organisms in hospital settings}},
  journal = {Npj Digit. Medicine},
  year = {2025},
  doi = {10.1038/s41746-025-01529-x},
}
Figure